MAPREDUCE Programming Model

Transcription

1 CS 2510 COMPUTER OPERATING SYSTEMS Cloud Computing MAPREDUCE Dr. Taieb Znati Computer Science Department University of Pittsburgh MAPREDUCE Programming Model Scaling Data Intensive Application MapReduce Framework Map Reduce Data Flow MapReduce Execution Model MapReduce Jobtrackers and Tasktrackers Input Data Splits Scheduling and Synchronization Speculative Execution Partitioners and Combiners Scaling Data Intensive Application Example WordCount Program I I can not do everything, but still I can do something; and because I cannot do everything, I will not refuse to do something I can do Word Count And 1 Because 1 But 1 Can 4 Do 4 Everything 2 I 5 Not 3 Refuse 1 Something 2 Still 1 To 1 Will 1 Define WordCount as Multiset; For Each Document in DocumentSet { T = tokenize(document); For Each Token in T { WordCount[token]++; } } Display(WordCount); Program Does NOT Scale for Large Number of Documents 1

2 WordCount Program II A two-phased program can be used to speed up execution by distributing the work over several machines and combining the outcome from each machine into the final word count Phase I Document Processing Each machine will process a fraction of the document set Phase II Count Aggregation Partial word counts from individual machines are combined into the final word count WordCount Program II Phase I Define WordCount as Multiset; For Each Document in DocumentSubset { T = tokenize(document); For Each Token in T { WordCount[token]++; } } SendToSecondPhase( wordcount); Phase II Define TotalWordCount as Multiset; For each WordCount Received from Phase I { MultisetAdd (TotalWordCount, WordCount); } WordCount Program II Limitations The program does not take into consideration the location of the documents Storage server can become a bottleneck, if enough bandwidth is not available Distribution of documents across multiple machines removes the central server bottleneck Storing WordCount and TotalWordCount in the memory is a flaw When processing large document sets, the number of unique words can exceed the RAM capacity In Phase II, the aggregation machine becomes the bottleneck WordCount Program Solution The aggregation phase must execute in a distributed fashion on a cluster of machines that can run independently To achieve this, functionalities must be added Store files over a cluster of processing machines. Design a disk-based hash table permitting processing without being limited by RAM capacity. Partition intermediate data across multiple machines Shuffle the partitions to the appropriate machines 2

3 Divide and Conquer Work Partition W 1 W 2 W 3 How to Scale Up? Worker Worker Worker R 1 R 2 R 3 Final Output Combine Parallelization Challenges How do we assign work units to workers? What if we have more work units than workers? What if workers need to share partial results? How do we aggregate partial results? How do we know all the workers have finished? What if workers fail? Parallelization Challenges Parallelization problems arise from in several ways Communication between workers, asynchronously Workers need to exchange information about their states Concurrent access to shared resources, while preserving state consistency Workers need to manipulate data, concurrently Cooperation requires synchronization and interprocess communication mechanisms 3

4 Distributed Workers Coordination Coordinating a large number of workers in a distributed environment is challenging The order in which workers run may be unknown The order in which workers interrupt each other may be unknown The order in which workers access shared data may be unknown Failures further compound the problem! Classic Models Computational Models Master and Slaves Producers and Consumers Readers and Writers IPC Models Shared Memory Threads Message Passing To ensure correct execution several mechanisms are needed Semaphores (lock, unlock), Conditional variables (wait, notify, broadcast), Barriers, Address deadlock, livelock, race conditions,... Makes it difficult to debug parallel execution on clusters of distribute processors MapReduce Data-Intensive Programming Model MapReduce is a programming model for processing large sets Users specify the computation in terms of a map() and a reduce() function, Underlying runtime system automatically parallelizes the computation across large-scale clusters of machines, and Underlying system also handles machine failures, efficient communications, and performance issues. MapReduce is inspired by the map() and fold() functions commonly used in functional programming Typical Large-Data Problem At a high-level of abstraction, MapReduce codifies a generic recipe for processing large data set Iterate over a large number of records Extract something of interest from each Map Shuffle and sort intermediate results Aggregate intermediate results Reduce Generate final output Basic Tenet of MapReduce is Enabling a Functional Abstraction for the Map() and Reduce() operations (Dean and Ghemawat, OSDI 2004) 4

5 MAPREDUCE Functional Programming Paradigm Imperative Languages and Functional Languages The design of the imperative languages is based directly on the von Neumann architecture Efficiency is the primary concern, rather than the suitability of the language for software development The design of the functional languages is based on mathematical functions A solid theoretical basis that is also closer to the user, but relatively unconcerned with the architecture of the machines on which programs will run Fundamentals of Functional Programming Languages The basic process of computation is fundamentally different in a FPL than in an imperative language In an imperative language, operations are done and the results are stored in variables for later use Management of variables is a constant concern and source of complexity for imperative programming FPL takes a mathematical approach to the concept of a variable Variables are bound to values, not memory locations A variable s value cannot change, which eliminates assignment as a possible operation Characteristics of Pure FPLs Pure FP languages tend to Have no side-effects Have no assignment statements Often have no variables! Be built on a small, concise framework Have a simple, uniform syntax Be implemented via interpreters rather than compilers Be mathematically easier to handle 5

6 Importance of FP FPLs encourage thinking at higher levels of abstraction It enables programmers to work in units larger than statements of conventional languages FPLs provide a paradigm for parallel computing Absence of assignment provide basis for independence of evaluation order Ability to operate on entire data structures FPL and IPL Example Summing the integers 1 to 10 in IPL The computation method is variable assignment total = 0; for (i = 1; i 10; ++i) total = total+i; Summing the integers 1 to 10 in FPL The computation method is function application. sum [1..10] 22 Lambda Calculus The lambda calculus is a formal mathematical system to investigate functions, function application and recursion. A lambda expression specifies the parameter(s) and the mapping of a function in the following form λ x. x * x * x for the function cube (x) = x * x * x Lambda expressions describe nameless functions Lambda expressions are applied to parameter(s) by placing the parameter(s) after the expression (λ x. x * x * x) 3 => 3*3*3 => 27 (λ x,y. (x-y)*(y-x)) (3,5) => (3-5)*(5-3) => -4 MAPREDUCE Functional Programming 6

7 FPL Map and Fold map and fold FPL higher-order functions (map f list1 [list2 list3 ]) (map square ( )) ( ) (fold f list [ ]) (fold + ( )) 30 (fold + (map square (map list1 list2)))) Roots in Functional Programming Map Fold f f f f f g g g g g What is MapReduce? Programming model for expressing distributed computations at a massive scale Execution framework for organizing and performing such computations Open-source implementation called Hadoop Mappers And Reducers A mapper is a function that takes as input one ordered (key; value) pair of binary strings. As output the mapper produces a finite multiset of new (key, value) pairs. Mappers operates on ONE (key; value) pair at a time A reducer is a function that takes as input a binary string k which is the key, and a sequence of values v 1, v 2,, v n, which are also binary strings. As output, the reducer produces a multiset of pairs of binary strings (k,v k,1 ), (k,v k,2 ), (k,v k,3 ), (k,v k,n ) Key in output tuples is identical to the key in input tuple. Consequence Mappers can manipulate keys arbitrarily, but Reducers cannot change the keys at all 7

8 MapReduce Framework Programmers specify two functions: map (k, v) <k, v >* reduce (k, v ) <k, v >* All values, v, with the same key are sent to the same reducer The execution framework supports a computational runtime environment to handle all issues related to coordinating the parallel execution of a data-intensive computation in a large-scale environment Breaking up the problem into smaller tasks, coordinating workers executions, aggregating intermediate results, dealing with failures and softeare errors, MapReduce Runtime Basic Functions Handles scheduling Assigns workers to map and reduce tasks Handles data distribution Moves processes to data, not data to processes Handles synchronization among workers Gathers, sorts, and shuffles intermediate data Handles errors and faults, dynamically Detects worker failures and restarts K 1 V 1 K 2 V 2 K 3 V 3 K 4 V 4 K 5 V 5 K 6 V 6 MapReduce Data Flow Input Records k 1 v 1 k 1 v 1 Output Records map map map map k 21 v 23 k 1 v 3 a 1 b 2 c 3 c 6 a 5 c 2 b 7 c 8 Shuffle and Sort: aggregate values by keys a 1 5 b 2 7 c Split k 12 v 32 Local QSort k 1 v 5 Reduce Reduce Reduce Split k 21 v 45 k 12 v 54 k 2 v 2 k 2 v 4 r 1 s 1 r 2 s 2 r 3 s 3 8

9 MapReduce Programmers specify two functions: map (k, v) <k, v >* reduce (k, v ) <k, v >* All values with the same key are reduced together The execution framework handles everything else! Not quite Usually, programmers also specify: partition (k, number of partitions) partition for k Often a simple hash of the key, e.g., hash(k ) mod N Divides up key space for parallel reduce operations combine (k, v ) <k, v >* Mini-reducers that run in memory after the map phase Used as an optimization to reduce network traffic MapReduce Design Issues Barrier between map and reduce phases To enhance performance the process of copying intermediate data can start early Keys arrive at each reducer in sorted order No enforced ordering across reducers MapReduce Implementations Google has a proprietary implementation in C++ Bindings in Java, Python Hadoop is an open-source implementation in Java Development led by Yahoo, used in production Now an Apache project Rapidly expanding software ecosystem Lots of custom research implementations For GPUs, cell processors, etc. Distributed Execution Overview Input data Split 0 Split 1 Split 2 Read Worker Worker Worker User Program Fork Fork Fork Assign Map Local Write Master Remote Read, Sort Assign Reduce Worker Worker Write Output File 0 Output File 1 9

10 Map-Reduce Framework for Word Count Input: a collection of keys and their values User- Written Map Function Each input (k,v) mapped to set of intermediate key-value pairs Sort All Key-value Pairs By Key Hello World : Word Count Each list of values is reduced to a single value for that key User- Written Reduce Function One list of intermediate values for each key: (k, [v 1,,v n ]) WordCount PseudoCode Map Reduce map(string filename, String document) { List<String> T = tokenize(document); for each token in T { emit ((String)token, (Integer) 1); } } reduce(string token, List<Integer> values) { Integer sum = 0; for each value in values { sum = sum + value; } emit ((String)token, (Integer) sum); } Implementation of WordCount() A program forks a master process and many worker processes. Input is partitioned into some number of splits. Worker processes are assigned either to perform Map on a split or Reduce for some set of intermediate keys. 10

11 Execution of WordCount Responsibility of the Master From Other Map Processes (w,1), (w,1), One Reduce Process (w,200), Note: typically one Reduce will handle many words. 1. Assign Map and Reduce tasks to Workers. 2. Check that no Worker has died (because its processor failed). 3. Communicate results of Map to the Reduce tasks. (w,1), (w,1), One Chunk of Docs One Map Process Distribute By Word (w,1), (x,1), (w,1), (y,1), Communication from Map to Reduce Select a number R of reduce tasks. Divide the intermediate keys into R groups, Use an efficient hashing function Each Map task creates, at its own processor, R files of intermediate key-value pairs, one for each Reduce task. MAP REDUCE Execution Framework Dr. Taieb Znati Computer Science Department University of Pittsburgh 11

12 MAPREDUCE Execution Framework MapReduce Execution Model MapReduce Jobtrackers and Tasktrackers Input Data Splits MapReduce Execution Issues Scheduling and Synchronization Speculative Execution Partitioners and Combiners MapReduce Data Flow A MapReduce job is a unit of work to be performed Job consists of the MapReduce Program, the Input data and the Configuration Information The MapReduce job is divided it into two types of tasks map tasks and reduce tasks It is not uncommon for MapReduce jobs to have thousands of individual tasks to be assigned to cluster nodes The Input data is divided into fixed-size pieces called splits One map task is created for each split The user-defined map function is run on each split Configuration information indicates where the input lies, and the output is stored Lifecycle of a MapReduce Job Map function Reduce function Job Execution Jobtrackers and Takstrackers Two types of nodes control the job execution process Jobtraccker and Tasktrackers A jobtracker coordinates all the jobs run on the system by scheduling tasks to run on Tasktrackers Tasktrackers run tasks and send progress reports to the jobtracker Jobtracker keeps a record of the overall progress of each job If a task fails, the jobtracker can reschedule it on a different Tasktracker 12

13 Scheduling and Synchronization Jobtracker Scheduling and Coordination In large jobs, the total number of tasks may exceed the number of tasks that can be run on the cluster concurrently, The Jobtracker must maintain a task queue and assign nodes to waiting tasks as the nodes become available. Another aspect of Jobtracker s responsibiliies involves coordination among tasks belonging to different jobs Jobs from different users, for example Designing a large-scale, shared resource to support several users simultaneously in a predictable, transparent and policy-driven fashion is challening! MapReduce Stragglers The speed of a MapReduce job is sensitive to the stragglers performance tasks that take an usually long time to complete The map phase of a job is only as fast as the slowest map task. The running time of the slowest reduce task determines the completion time of a job Stragglers may result from unreliable hardware A machine recovering from frequent hardware errors may become significantly slower The barrier between the map and reduce tasks further compounds the problem MapReduce Speculative Execution Speculative execution is an optimization technique to improve job running times, in the presence of stragglers Base on speculative execution, an identical copy of the same task is executed on a different machine, The result of the task that finishes first is used Google has reported that speculative execution can achieve 44% performance improvement 13

14 MapReduce Speculative Execution Both map and reduce tasks can be speculatively executed, but the technique is better suited to map tasks than reduce tasks This is due to the fact that each copy of the reduce task needs to pull data over the network. Speculative execution cannot adequately address cases where stragglers are caused by a skew in the distribution of values associated with intermediate keys In these cases, tasks responsible for processing the most frequent elements run much longer than the typical task More efficient local aggregation may be required MapReduce Synchronization In MapReduce, synchronization is needed when mappers and reduces exchange intermediate output and state information Intermediate key-value pairs must be grouped by key, which requires the execution a distributed sort process involving all the nodes that executed map tasks and all the nodes that will execute reduce tasks The shuffle and sort process involves copying intermediate data over the network MapReduce Synchronization A MapReduce job with M mappers and R reducers may involves up to M R distinct copy operations Each mapper intermediate output goes to every reducer No reducer can start until all the mappers have finished emitting key-value pairs and all intermediate key-value pairs have been shuffled and sorted Necessary to guarantee that all values associated with the same key have been gathered. This is an important departure from functional programming, where aggregation can begin as soon as values are available. For improvement start copying intermediate key-value pairs over the network to the nodes running the reducers as soon as each mapper finishes Data Locality Optimization 14

15 Input Division Split Size Fine-grained splits increase parallelism and improves fault-tolerance Small splits reduce the processing time of each split and allows faster machines to process proportionally more splits over the course of the job than slower machines Load-balancing can be achieved more efficiently with small splits The impact of failure, when combined with load-balancing, can be reduced significantly with fine-grained splits Too small splits increases the overhead of managing splits Map task creation dominates the total job execution time. Data Locality Input Data Data locality Optimization The map task should be run on a node where the input data resides The optimal split size is the same as the largest size of input that can be guaranteed to be stored on a single node. If the split is larger than what one node can store, data transfer on across the network to the node running the map task is required May result in significant communication overhead, and reduces efficiency Data Locality Map Output Output produced by map tasks should be stored locally, NOT at a distributed storage Map output is intermediate Processed by reduce tasks to produce the final output Map output is no needed upon completion of the job Map output should NOT be replicated to overcome failure It is more efficient to restart the map task upon failure than replicating the output produced by map tasks Data Locality Reduce Tasks Reduce tasks cannot typically take advantage of data locality Input to a single reduce task is normally the output from all mappers. The sorted map outputs have to be transferred across the network to the node where the reduce task is running The outputs are merged and then passed to the user-defined reduce function 15

16 Reduce Task Output Replication Multiple replicas of the reduce output are normally stored reliably, First replica is stored on the local node, Other replicas being stored on off-rack nodes. To increase efficiency, the writing of the reduce output must reduce the amount of network bandwidth consumed The replication process must be streamlined and efficient Partitioners and Combiners Number of Reduce Tasks The number of map reducers is typically specified independently It is not only governed by the size of the input, but also the type of the application MapReduce data flow can be specified in different ways No reduce tasks Singe reduce task Multiple reduce tasks Data Flow No Reduce Tasks Input Split 0 Map Output Part o Split 1 Map Part 1 Split 2 Map Part 2 Replication Replication Replication 16

17 Data Flow Single Reduce Task Input Split 0 Map Sort Copy Output Merge Split 1 Map Sort Reduce Part 0 Replication Data Flow Multiple Reduce Tasks Input Sort Split 0 Map Split 1 Map Output Copy Merge Reduce Part 0 Replication Merge Reduce Part 1 Replication Split 1 Map Sort Split 1 Map Data Flow Multiple Reduce Tasks Map tasks partition their output, each creating one partition for each reduce task There can be many keys and their associate values in each partition The recodes for every key are all in a single partition The partition can be controlled by user-defined partition function The use of a hash function typically works well The Shuffle, data flow between map and reduce, is a complicated process whose tuning can have a big impact on the job execution MapReduce Combiners Combiner functions can be used to minimize the data transferred between map and reduce tasks Combiners are particularly useful when MapReduce jobs are limited by the bandwidth available on the cluster Combiners are user-specified functions Combiner functions run on the map output The combiner s is then fed to the reduce function Since it is an optimization function, there is no guarantee how many times combiners are called for a particular map output record, if at all Calling the combiner zero, one, or many times should produce the same output from the reducer. 17

18 Combiner Example Max Temperature Assume that the mappers produce the following output Mapper 1 (1950, 0), (1950, 20), (1950, 10) Mapper 2 (1950, 25), (1950, 15) Reduce function is called with a list of all the values (1950, [0, 20, 10, 25, 15]) with output (1950, 25) Using a combiner for each map output results in: Combiner 1 (1950, 20) Combiner 2 (1950, 25) Reduce function is called with (1950, [20,25]) with output (1950, 25) Combiner Property The combiner function calls can be expressed as follows: Max(0, 20, 10, 25, 15) = Max(Max(0, 20, 10), Max(25, 15)) = Max(20, 25)=25 Max() is commonly referred to as distributive Not all function exhibit distributive property Mean(0, 20, 10, 25, 15) = ( )/5=14 Mean(Mean(0, 20, 10), Mean(25, 15)) = Mean (10, 20) = 15 Combiners do not replace producers Producers are still needed to process recorders with the same key from different maps Conclusion Part I Conclusion Part II MapReduce Execution Framework MapReduce Jobtrackers and Tasktrackers Input Data Splits MapReduce Execution Issues Scheduling and Synchronization Speculative Execution Partitioners and Combiners Scaling Data Intensive Application MapReduce Framework MapReduce Overview Map Reduce Data Flow Map Function Partition Function Compare Function Reduce Function 18

19 Reference Data-Intensive Text Processing with MapReduce, Jimmy Lin and Chris Dyer. MapReduce: Simplified Data Processing on Large Clusters, Jeffrey Dean and Sanjay Ghemawat, The Google File System, Sanjay Ghemawat, Howard Gobioff, and Shun-TakLeung, 19